Customer Support

How to process refund and return requests with AI

Customer Support3 AI tools7 steps6 friction points

Want an agent to handle this on your live data? See the Starch version →

Processing refund and return requests is one of those workflows that looks simple on the surface — a customer wants their money back, you decide whether to approve it — but in practice it involves pulling order history, checking return windows, applying policy rules, updating your payment processor, logging the resolution, and communicating back to the customer. Do that ten times a day and it's a real operational load. Do it inconsistently and you erode trust.

It feels like an AI problem because most of the work is pattern-matching against a known policy. Given an order date, a return window, and a reason code, the right answer is usually deterministic. The actual judgment calls — fraud signals, edge cases, goodwill exceptions — are a small fraction of the volume. Operators reach for ChatGPT or Claude hoping the AI can just handle the routine 80% so they can focus on the exceptions.

General-purpose AI tools like ChatGPT, Claude, and Gemini can genuinely help here. They can parse return requests, apply your stated policy, draft denial or approval emails, and summarize edge cases for human review. The limitation isn't the reasoning — it's the data access. The LLM can only work with what you paste into the conversation, and it has no connection to your actual orders, payment processor, or CRM.

Skip the manual work → See the Starch version

Customer Support3 AI tools7 steps6 friction points

AI walkthrough

How to do it with AI today

A practical walkthrough using ChatGPT, Claude, and other off-the-shelf LLMs — what they're good at, what you'll have to do by hand.

Tools that work for this

ChatGPTClaudeGemini

Step-by-step

1 Export the relevant order data from your e-commerce platform (Shopify, WooCommerce, etc.) or payment processor as a CSV — at minimum: order ID, purchase date, customer email, item(s), order value, and any previous refund history.

2 Paste your return policy into the LLM as a system prompt or top-of-conversation context. Be explicit: return window in days, eligible versus ineligible item categories, conditions for full versus partial refund, who pays return shipping.

3 For each incoming request, paste the customer's message alongside the relevant order row from your export. Ask the LLM to evaluate whether the request falls within policy, flag any ambiguity, and recommend approve, deny, or escalate.

4 Use the LLM to draft a response email for each decision. Give it a tone guide — how formal, whether to include a return label link placeholder, how to phrase a denial without triggering a chargeback.

5 For approvals, manually trigger the refund in your payment processor (Stripe, Square, PayPal) using the order ID the LLM surfaced. The LLM cannot do this — it only produces the decision and the draft.

6 Log the outcome manually in your CRM or support tracker. Paste the customer name, order ID, decision, reason, and refund amount into wherever you track support history.

7 At the end of each week, paste all resolved requests back into the LLM and ask it to summarize patterns — top return reasons, refund rate by product category, any policy gaps the volume reveals.

Prompts you can copy

Here is my return policy: [paste policy]. Here is the customer's request: [paste message]. Here is their order data: [paste row]. Should I approve, deny, or escalate? Explain your reasoning in one sentence.

Draft a polite denial email for a return request that is outside the 30-day window. The customer says the item arrived damaged. Acknowledge the frustration, explain the policy, and offer a 15% discount on their next order as a goodwill gesture.

Here are 12 resolved return requests from this week: [paste data]. Summarize the top 3 return reasons, identify any products appearing more than once, and flag any requests where our policy decision seems inconsistent.

The customer is requesting a refund on order #4821, placed 45 days ago. Our policy allows 30 days. They claim they were traveling. Write two response options: one that holds the policy and one that approves as a one-time exception, and note the tradeoffs of each.

I am building a return request triage rubric. Given these return reason categories — wrong item, damaged in shipping, changed mind, defective after use, wrong size — what policy rules should I apply to each, and what information should I always collect before deciding?

Reality check

Where this gets hard

The walkthrough above works — until your numbers change, the LLM hallucinates, or you have to re-paste everything next month.

No live connection to your Shopify or Stripe data — every session starts with a manual export and paste, which means the LLM is always working off a snapshot, not the actual current order.

Volume doesn't scale. Pasting orders one at a time is fine for five requests; it breaks down at fifty. Batch pasting hits context limits and the LLM starts dropping rows or losing track of which response belongs to which order.

The LLM can draft the approval email but it cannot actually issue the refund, update the order status, or log the resolution in your CRM. Every decision still requires manual action in two or three other tools.

Nothing persists between sessions. The policy context you carefully formatted last Tuesday has to be re-pasted today. If your policy changes, there is no single source of truth — just whatever is in the most recent prompt.

Tone and format drift. The denial email the LLM wrote last week had a specific structure your team agreed on. This week it writes something slightly different. Enforcing consistency requires constant re-prompting.

No audit trail. If a customer disputes a decision or a chargeback is filed, you have no record of why the AI recommended what it did — the conversation is gone unless you manually copy it somewhere.

Tired of the friction?

Starch runs the whole workflow on live data — no copy-paste, no hallucinated numbers, no re-prompting next month.

See the Starch version →

Starch alternative

The same workflow on Starch

Starch is an agentic operating system — it builds and runs the software that handles this workflow continuously, connected to your live data, instead of a one-off prompt you re-run by hand every time a request lands in your inbox.

Connect Shopify or your payment processor from Starch's integration catalog — the agent queries live order data when a return request comes in, so decisions are based on the actual order, not a CSV you exported three days ago.

Starch syncs your Gmail or Outlook on a schedule. Describe what a return request looks like and the agent routes incoming requests automatically — no manual triage, no inbox monitoring required.

The CRM starter app tracks every return interaction by customer — approval history, reason codes, refund amounts, follow-up status — so you always know if a customer is a repeat requester before you decide on an exception.

Describe your refund policy in plain English and Starch builds an approval app around it: the agent evaluates each request against your rules, drafts the customer-facing response, and surfaces only the edge cases that need your judgment.

Customer Support Agent — coming soon — will handle return and refund tickets end-to-end across email and chat, using your policy and order data as the source of truth, and escalating to your team with full context when a case needs a human call.

Automations trigger on events, not on your memory. Tell Starch: 'When a refund is approved, log it to the CRM, send the customer a confirmation, and Slack me a daily summary of refund volume by product.' It builds and runs that chain without you re-assembling it each time.

Get closed-beta access →

Toolkit